I'm looking to build an object detection model or image segmentation model, ideally the latter, which will identify and label objects from satellite imagery but I don't currently have any labelled data.
I have looked into using existing labelled satellite data but most do not have the variety of objects I desire, also I think I'd like to put together my own set for the experience. I am going to create a separate post on Open Data Exchange regarding this, but if anyone does has some notable labelled satellite images I'd still be happy to check them out.
I am now looking at the best/most efficient way to build up my own labelled data set. I'd like to automate this process as much as possible to build a suitable size collection of potential images to be labelled. These images can be fed into a system where users can select what the item is, or whether to discard the image. This will then build up a labelled data set which can be further used with ML to detect objects / image pixel segments and classify/label them.
So far the best I can seem to find is using some sort of semi-supervised / weak supervision ML models that still initially require some sort of labelled data set. I would consider this an option, but to do so I'd need to know how many labelled images I'd need put together for each object before taking that route. How much labelled data on average would be required per item to detect for semi-supervised / weak supervision object detection/segmentation?
Preferably I'd like to be able to start without any labelled data and have something output a set of unlabelled potential objects which can then be manually discarded because its not an object or labelled as an object and what type of object it is. This labelled dataset could then be further used as input for other ML models detect and label/classify. What, if any, options are there for creating unlabelled/unclassified object detection or image segmentation without any initial input labelled data?